• 综合
  • 标题
  • 关键词
  • 摘要
  • 学者
  • 期刊-刊名
  • 期刊-ISSN
  • 会议名称
搜索

作者:

Fu, Pengbin (Fu, Pengbin.) | Xiao, Ganyun (Xiao, Ganyun.) | Yang, Huirong (Yang, Huirong.)

收录:

Scopus SCIE

摘要:

The complex two-dimensional structure poses huge challenges for handwritten mathematical expression recognition (HMER). Many researchers process the LaTeX sequence into a tree structure and then design tree decoders based on RNN to address this issue. However, RNNs have problems with long-term dependency due to their structural characteristics. Although Transformers solve the long-term dependency problem, tree decoders based on Transformers are rarely used for HMER because the attention coverage is significantly insufficient when the distance between parent and child nodes is large in tree structures. In this paper, we propose a novel offline HMER model SATD incorporating a tree decoder based on Transformer to learn the implicit structural relationships in LaTeX strings. Moreover, to address the issue of distant parent-child nodes, we introduce a multi-scale attention aggregation module to refine attention weights using contextual information with different receptive fields. Experiments on CROHME2014/2016/2019 and HME100K datasets demonstrate performance improvements, achieving accuracy rates of 63.45%/60.42%/61.05% on the CROHME 2014/2016/2019 test sets. The source code https://github.com/EnderXiao/SATD/ of this work will be publicly available.

关键词:

Offline handwritten mathematical expression recognition Attention Coverage attention Tree decoder Transformer

作者机构:

  • [ 1 ] [Fu, Pengbin]Beijing Univ Technol, Fac Informat Technol, Xidawang Rd, Beijing 100124, Peoples R China
  • [ 2 ] [Xiao, Ganyun]Beijing Univ Technol, Fac Informat Technol, Xidawang Rd, Beijing 100124, Peoples R China
  • [ 3 ] [Yang, Huirong]Beijing Univ Technol, Fac Informat Technol, Xidawang Rd, Beijing 100124, Peoples R China

通讯作者信息:

  • [Yang, Huirong]Beijing Univ Technol, Fac Informat Technol, Xidawang Rd, Beijing 100124, Peoples R China;;

查看成果更多字段

相关关键词:

相关文章:

来源 :

VISUAL COMPUTER

ISSN: 0178-2789

年份: 2024

3 . 5 0 0

JCR@2022

被引次数:

WoS核心集被引频次:

SCOPUS被引频次:

ESI高被引论文在榜: 0 展开所有

万方被引频次:

中文被引频次:

近30日浏览量: 0

归属院系:

在线人数/总访问数:321/4977230
地址:北京工业大学图书馆(北京市朝阳区平乐园100号 邮编:100124) 联系我们:010-67392185
版权所有:北京工业大学图书馆 站点建设与维护:北京爱琴海乐之技术有限公司